Classification of personnel and vehicle activity using a sensor system with numerous array elements

There is increasing interest in the use of sensors with a large number of elements for persistent coverage over large areas in terrestrial applications. In support of this, algorithms for detection and classification of footsteps, pounding activity, and vehicle activity are being developed. In practice, many applications take place in a busy environment, leaving an operator overwhelmed investigating every detection. In such cases, robust automatic classification becomes of primary importance. Such applications offer the additional challenge of classifying a large and diverse set of signals of interest, subject to the ¿curse of dimensionality¿ brought about by estimating probability density functions in a common, high-dimensional feature space. Using a Class-Specific Classifier offers the advantage of allowing separate low-dimensional feature sets for each class. In this paper, a detailed description of the signals of interest, detector, and classifier are presented. The performance of a hybrid discriminative/generative classifier is presented using experimental data collected from a scripted field test. Results demonstrate classifier performance of over 90% probability of correct classification for all classes of interest.